Power efficient routing trees for ad hoc wireless networks using directional antenna

Ad Hoc Networks 3 (2005) 621 628 www.elsevier.com/locate/adhoc Power efficient routing trees for ad hoc wireless networks using directional antenna Fei Dai a, *, Qing Dai a, Jie Wu a a Department of Computer Science and Engineering, Florida Atlantic University, Boca Raton, FL 33431, USA Available online 15 September 2004 Abstract In ad hoc wireless networks, nodes are typically powered by batteries. Therefore saving energy has become a very important objective, and different algorithms have been proposed to achieve power efficiency during the routing process. Directional antenna has been used to further decrease transmission energy as well as to reduce interference. In this paper, we discuss five algorithms for routing tree construction that take advantage of directional antenna, i.e., Reverse- Cone-Pairwise (), Simple-Linear (SL), Linear-Insertion (), Linear-Insertion-Pairwise (P), and a traditional approximation algorithm for the travelling salesman problem (TSP). Their performances are compared through a simulation study. Ó 2004 Elsevier B.V. All rights reserved. Keywords: Ad hoc wireless network; Directional antenna; Energy-efficient routing; Transmission power; Travelling salesman problem (TSP) 1. Introduction: Background and related work Energy efficiency is an important consideration for ad hoc wireless networks, where nodes are typically powered by batteries. It is directly correlated with network longevity and connectivity, therefore affecting network throughput. Among all different components of power consumption, transmission cost appears to dominate, compared to receiving * Corresponding author. E-mail addresses: fdai@fau.edu (F. Dai), qdai_jfgao@hotmail.com (Q. Dai), jie@cse.fau.edu (J. Wu). cost and computation cost. It has been shown that the power threshold p for a source node to reach its destination node is positively correlated to the distance between them, and can usually be expressed as p = r a + c for some constants a and c [1], where r is the distance between the two nodes, and a is between 2 and 4. In previous studies, different metrics have been used. Some measure the overall energy consumption of the network, while some others try to extend lifespan of individual nodes. The Broadcast Incremental Power algorithm (BIP) is a centralized algorithm attempting to minimize the overall energy in route 1570-8705/$ - see front matter Ó 2004 Elsevier B.V. All rights reserved. doi:10.1016/j.adhoc.2004.08.008

622 F. Dai et al. / Ad Hoc Networks 3 (2005) 621 628 determination [2]. It is similar to PrimÕs Minimum Spanning Tree (MST) algorithm [3], in that at any time all reached nodes form a single-rooted tree. Each step adds the node with the minimum incremental cost, calculated as the additional energy for it to be reached by any node within the tree, either by increasing power of a transmitting node, or by making a non-transmitting node transmit at a specific power level. It has already been proved that BIP has a constant approximation ratio of between 6 and 12, compared to the optimal solution [4]. On the other hand, some power-aware routing approaches select routes that avoid nodes with low remaining power, which can be either absolute or relative power level. Different metrics may well lead to different algorithms, but they can also be combined to achieve a balance, thus optimizing overall energy consumption without depleting crucial nodes [5,6]. Recently, the use of directional antenna was proposed in order to reduce interference and to further increase power efficiency, because ideally, power consumed in the case of directional antenna is only r a h 2p ; where h is the beam angle. In one of the sessionbased studies, two algorithms were proposed: RB-BIP (Reduce Beam BIP) and D-BIP (Directional BIP). The former simply adds a beamreducing step to the original BIP algorithm, so that each node can now transmit at its smallest possible angle. The latter incorporates the use of directional antenna at each step of the tree construction, i.e., a node in the tree could also increase its current transmission beam angle or shift the existing beam to reach a new destination node, in addition to increasing its transmission power, whichever gives the lowest incremental cost [5]. In our study, we try to find different centralized routing algorithms for broadcasting with the use of directional antenna. Four new algorithms are proposed to create a routing tree in ad hoc networks. In the Reverse-Cone-Pairwise () algorithm, the transmission beam of each sender is adjustable to achieve a good balance between the number of covered receivers and the transmission power. This algorithm is a refinement to the original D-BIP scheme. The difference is that employs a new transmission beam adjustment scheme, which incurs less incremental cost when expanding the beam to cover an additional receiver. The remaining three algorithms, Simple- Linear (), Linear-Insertion (), and Linear- Insertion-Pairwise (P), form a linear chain to connect all nodes in the network, where each sender uses the minimal beam angle to reach its next hop in the chain. These algorithms have lower computation cost than, and is appropriate for ad hoc networks using very narrow transmission beams. We also investigate a traditional travelling salesman (TSP) algorithm, which can also form a chain for broadcasting. All those schemes are evaluated via a simulation study, and their energy efficiency is compared with that of D-BIP. The remainder of this paper is organized as follows: five algorithms are introduced in Section 2, including four new algorithms and one existing algorithm for comparison. We then simulate their performance with randomly generated networks in Section 3, and discuss their applications and potentials in Section 4. 2. Algorithms Our objective is to find routing algorithms that are power efficient using directional antenna. In this study, we assume one beam for each node, and a = 2 and c = 0 to calculate the transmission energy. Here we propose four algorithms, assuming global knowledge of node locations, adjustable transmission range and adjustable beam angle. In all four algorithms, we placed a constraint that each sender forms only one transmission beam to reach its receivers. Simultaneously forming several beams is possible, but requires extra hardware support. Multiple beams can be emulated via consequent transmissions in several directions. However, switching transmission directions causes energy and delay penalty. Therefore, it is better to keep the beam direction as long as possible, instead of changing direction frequently [7]. In addition, allowing multiple beams renders the original problem a trivial one, which is equivalent to con-

F. Dai et al. / Ad Hoc Networks 3 (2005) 621 628 623 structing a MST when the minimum beam angle is very small. The first algorithm is a refinement of D-BIP [5], where the node with the minimum incremental cost is added at each step, while taking the adjustable antenna beam width into consideration (see Fig. 1). The actual beam angle that is used has to exceed a minimum beam angle minangle. All nodes that have already been reached can act as transmitting node, and all the non-tree nodes are potential destination nodes. Each possible transmitting-destination node pair is examined to determine the minimum power and beam angle needed to add a new destination node. When a transmitting node is adding a new destination node, if the new node does not fall into its current transmission beam, it can either shift or expand its previous beam to cover the new node. The incremental power is determined by subtracting the energy of the previous transmission beam from the new power, calculated as incrementalcostði; jþ ¼ðp 0 h 0 p hþ=ð2 pþ; where p 0 and h 0 represent the new transmission power and beam angle, respectively, while p and h are the previous power and beam angle. This process is demonstrated in Fig. 2, where i and j are the source and destination nodes, respectively, and the cone represents the current transmission beam. In case 1, the destination node d could be enclosed within the current beam, including through a beam shift. It will be covered without Fig. 2. Incremental transmission cost. any incremental cost. Otherwise, either the transmission beam has to be expanded (case 2) or transmission power needs to be increased (case 3), or both (case 4) for d to be reached. It could therefore be very costly. During this process, it is desirable to use the minimum transmission beam angle whenever possible. A simple heuristic method to calculate the new beam is to keep the start and end point of the previous beam, and to expand either end that leads to a smaller increase of the overall beam span. We will see that this heuristic does not always provide the optimal beam angle. An example is illustrated in Fig. 3. At first, the source node a is transmitting to two downstream neighbors b and c with a beam angle h 1. When a new destination node d joins, and when h 1 is close to p, neither expanding beam to b c d nor to c b d provides the minimum beam. On the other hand, b d c is the optimal solution in this case with a beam angle h 2. The Reverse-Cone method (Fig. 4) is developed to calculate the minimum beam span. Instead of Fig. 1. Algorithm I: Reverse-Cone-Pairwise (). Fig. 3. Illustration of Reverse-Cone method.

624 F. Dai et al. / Ad Hoc Networks 3 (2005) 621 628 Fig. 4. Reverse-Cone method to find the minimum beam. merely keeping the start and end point of the previous beam, each node keeps angle positions, calculated as the radius of a destination node from itself, of all its destination nodes in a sorted list anglelist. The 2p circle around a transmitting node is then divided into several cones, each defined by the two adjacent neighbors in the anglelist. Whenever adding a new destination node, the transmitting node will first insert the angle location of the new node into its anglelist. The angle- List is traversed to find the largest cone. The minimum new beam to cover all the nodes in anglelist would be the reverse of this largest cone. For the previous example in Fig. 3, node a first adds nodes b and c in its sorted anglelist. Of the two cones, c b counterclockwise is larger than b c, therefore the reverse of cone c b, i.e., cone b c,orh 1, is selected as the minimum beam. When node a adds a new destination d, d is then inserted into aõs angle list, which now becomes d b c. After traversal of the new anglelist, b c is found to be the largest cone, therefore the reverse of it, c d b, orh 2 will be used as the new transmission beam. The algorithm generally provides good performance, except in the situation when the previous transmission beam was small, adding a new node at a distant angle position can be very expensive. Three other heuristic algorithms are therefore developed: Simple-Linear (SL), Linear-Insertion (), and Linear-Insertion-Pairwise (P). In all of them, each transmitting node always uses the minimum beam angle minangle to reach exactly one downstream node, and there is only one downstream destination node for each transmitting node. As a result, a linear chain will be formed step by step, starting from the source node. Initially, the reachednodes set includes sourcenode only, and the unreachednodes set includes all other nodes. In SL (see Fig. 5), initially the source node is the listhead, then at each step, a minnode is determined as an unreached node closest to the current listhead, and is added to the reachednodes set to become the new listhead. This listhead is the only possible transmitting node to reach the next new node, until the unreachednodes set is empty. In (Fig. 6), an additional backtrack step is included when each node is added to the reached- Nodes set. During the backtrack, after the min- Node is determined, it is first tested for possible insertions into each position between any two adjacent nodes within the existing linear chain. The insertioncost, the incremental cost for inserting the new node, is calculated to check whether any insertion causes a saving of energy, compared to directly adding the minnode as the new list- Head. If so, an insertion will take place where the insertioncost is the minimum. In this case, the previous listhead remains unchanged. If not, minnode would be attached as the new listhead. In the example in Fig. 7, s represents the source- Node, d 1 through d 6 are the destination nodes to be reached. The Linear-Insertion algorithm would add d 6 between s and d 1 in the backtrack process, which leads to a lower overall energy cost. While seems to be a better solution than SL, unfortunately, it does not always outperform SL due to its heuristic nature (examples not shown). P (Fig. 8) takes one further step beyond. In P, the incremental cost for inserting every Fig. 5. Algorithm II: Simple-Linear (SL).

F. Dai et al. / Ad Hoc Networks 3 (2005) 621 628 625 Fig. 8. Algorithm IV: Linear-Insertion-Pairwise (P). Fig. 6. Algorithm III: Linear-Insertion (). salesman problem (TSP), which is NP-complete. TSP has an approximation solution in geographic graphs [8], where each node is placed in a 2-D area, and the cost of an edge (i,j) is proportional to the distance of the two end nodes i and j. The TSP approximation algorithm (Fig. 9) consists of two phases. The first phase (lines 1 11) constructs a MST using PrimÕs algorithm [3]. The second phase (lines 13 21) constructs a chain by taking a preorder walk of the MST. This algorithm has an approximation ratio of 2. That is, the total cost of the path is at most twice that in an optimal solution. Although this approximation ratio does not Fig. 7. Illustration of Linear-Insertion. node in the unreachednode set to the existing chain is computed. The minnode with the minimum incremental cost, instead of the minimum distance to the listhead, is removed from the unreached- Node set and inserted into the reachednode chain. To simplify the description, we view attaching a node j to the listhead and make j the new listhead as a special case of insertion, and use Cost(list- Head,j) as the incremental cost in this case. In each single step, P can find a minnode with lower mincost than. Again, it is a heuristic approach, and does not always outperform. The problem of forming a minimum path to visit all nodes has been known as the travelling Fig. 9. Algorithm IV: Travelling Salesman Approximation.

626 F. Dai et al. / Ad Hoc Networks 3 (2005) 621 628 apply in non-geographic graphs, we can still use it as a benchmark in performance comparisons. The complexity of the five algorithms are as follows: and have the similar computation complexity of H(n 2 ). P and TSP have a higher complexity of H(n 3 ). has the highest complexity of H(n 4 ), with the Reverse-Cone method. To evaluate the relative performance of the five algorithms, simulations are performed on random networks, and their results are shown in the next section. 3. Simulation results We generate random network instances, and the four algorithms are used to obtain routing graphs, respectively. Their transmission costs are calculated and compared with each other. Our results show that in most occasions, produces routing solutions with good energy-efficiency. Between SL,, and P, has a better performance than SL, and P is slightly better than. In simulation with different network density, we observed that when node distribution is sparse, and P outperform when the minimal beam angle is relatively small. In the example shown in Fig. 10, we set the number of nodes from 10 to 100 in the same size of area. As the minimal beam angle increases gradually, we can see that energy costs of SL,, and P routing increase with the minimal angle linearly. This is because in those algorithms, the route construction process is not affected by the minimal angle, and the same routing graph will always be produced just as omnidirectional antenna is provided. The final cost is simply calculated by multiplying the original overall cost by a factor of minangle/2p. The result of is close to when the minimal angle is small, which makes sense since in this case would most likely generate the same routing graph as SL. When the minimal angle increases to a certain point, will outperform SL, and eventually outperform and P, when shifting or expanding current beam shows advantage. The difference between and P is trivial in sparse networks with 10 or 30 nodes. This difference becomes obvious in dense networks, where P can find a better min- Node in each step among a large pool of candidates. Simulations are also performed to measure the overall cost over the number of nodes. In a fixed area, the costs of SL, and P increment along with the increase of nodes, but the cost of decreases, except when the node number and minimal beam angle are very small (Fig. 11). It seems that in most cases, has the best scalability among the four. To compare the performance of our four algorithms, especially, with that of D-BIP, a simplified version of D-BIP is implemented. As before, we consider only the transmission cost, which is a function of the distance. As expected, shows better energy efficiency than D-BIP (Fig. 12). However, the difference among and D-BIP is very small. The same figure also shows the performance of the traditional TSP algorithm. It is obvious that all algorithms proposed in this paper outperform TSP. 1400 400 200 P 0 5 10 15 20 25 30 35 40 45 Minimum Beam Angle (degree) 1 1 P 1400 400 200 0 5 10 15 20 25 30 35 40 45 Minimum Beam Angle (degree) (a) (b) (c) 2500 2000 1500 500 P 0 5 10 15 20 25 30 35 40 45 Minimum Beam Angle (degree) Fig. 10. Simulation results with the minimum beam angle varying from p/36 (5 ) top/4 (45 ). (a) 10 nodes, (b) 30 nodes and (c) 100 nodes.

F. Dai et al. / Ad Hoc Networks 3 (2005) 621 628 627 750 700 650 550 500 450 400 P 350 10 20 30 40 50 60 70 80 90 100 Node Number 500 10 20 30 40 50 60 70 80 90 100 Node Number (a) (b) (c) 1500 1400 1300 1100 900 700 P 10 20 30 40 50 60 70 80 90 100 Node Number Fig. 11. Simulation results with number of nodes from 10 to 100. Minimal beam angle: (a) p/12 (15 ), (b) p/6 (30 ) and (c) p/4 (45 ). 2400 2200 2000 1 1 1400 P 2500 2000 2000 1500 TSP P D-BIP 1 1 1400 TSP P D-BIP 500 (a) 0 5 10 15 20 25 30 35 40 45 Minimum Beam Angle (degree) (b) 10 20 30 40 50 60 70 80 90 100 Node Number Fig. 12. Simulation results of proposed schemes compared with TSP and D-BIP. (a) 30 nodes and (b) minimal beam angle p/6 (30 ). Nevertheless, all four algorithms described above are greedy algorithms and heuristic in nature, with none being optimal. In fact, we can find special network instances between,, P or, where one outperforms the other three (examples not shown). From the simulation result, it seems that when nodes are relatively sparse and the minimal beam angle is small, is the best choice. Not only does it give the lowest routing cost, but its complexity is lower than also. The routing cost of P is close to that of, but P has a higher computation complexity than. When the minimal angle increases to a certain extent or when the network is dense, expanding current beam to include additional destination nodes would be more cost efficient, and at this point, provides the best performance. 4. Conclusion The problem of broadcasting a message to the entire ad hoc network with minimum energy is NP-complete. Using directional antennas can further reduce the energy consumption, but also makes the problem more complicated. In this paper, we have proposed four heuristic algorithms that construct a broadcast tree based on global information. Those algorithms have different performance in terms of reducing the broadcast cost and incur different computation cost. A simulation study has been conducted to provide guidelines for tradeoffs under different network situations. These algorithms are also compared with two existing algorithms in terms of their energy efficiency. Currently, we are further analyzing their performance statistically. In the future, it will be interesting to know the approximation ratio of the those algorithms. Acknowledgments The work was supported in part by a grant from Motorola Inc. and NSF grants CNS 0434533, CCR 0329741, ANI 0073736, and EIA 0139806.

628 F. Dai et al. / Ad Hoc Networks 3 (2005) 621 628 References [1] V. Rodoplu, T. Meng, Minimum energy mobile wireless networks, IEEE Journal on Selected Areas in Communications 17 (8) (1999) 1333 1344. [2] J.E. Wieselthier, G.D. Nguyen, A. Ephremides, On the construction of energy-efficient broadcast and multicast trees in wireless networks, in: Proceedings of IEEE INFO- COM, 2000, pp. 585 594. [3] R.C. Prim, Shortest connection networks and some generalizations, Bell System Technical Journal 36 (1957) 1389 1401. [4] P.J. Wan, G. Calinescu, S.Y. Li, O. Frieder, Minimumenergy broadcasting in static ad hoc wireless networks, Wireless Networks 8 (2002) 607 617. [5] J.E. Wieselthier, G.D. Nguyen, A. Ephremides, Energylimited wireless networking with directional antennas: the case of session-based multicasting, in: Proceedings of IEEE INFOCOM, 2002, pp. 190 199. [6] J. Wu, F. Dai, M. Gao, I. Stojmenovic, On calculating power-aware connected dominating set for efficient routing in ad hoc wireless networks, IEEE/KICS Journal of Communications and Networks 4 (2002) 59 70. [7] A. Spyropoulos, C.S. Raghavendra, Energy efficient communications in ad hoc networks using directional antennas, in: Proceedings of IEEE INFOCOM, 2002. [8] T.H. Cormen, C.E. Leiserson, R.L. Rivest, C. Stein, Introduction to Algorithms, 2nd ed., The MIT Press, Cambridge, MA, 2001. Fei Dai is a Ph.D. candidate at Department of Computer Science and Engineering, Florida Atlantic University. His current research focuses on the design and application of localized algorithms in wireless ad hoc and sensor networks. Other research areas include interconnection networks, fault tolerant routing, and software architecture. He received his B.S. and M.S. degrees in Computer Science in 1990 and 1993, respectively, from Nanjing University, Nanjing, China. He worked as a senior programmer in Greatwall Computer, China, from 1993 to 1996, and as a software architect and team leader in J&A Securities, China from 1996 to 2000. Qing Dai received her M.S. degree in Computer Science from Florida Atlantic University on August 2003, and M.S. degree in Microbiology from Upstate University on July 2000. She is currently a software engineer at Motorola, Plantation, FL. Jie Wu is a Professor at Department of Computer Science and Engineering, Florida Atlantic University. He has published over 200 papers in various journal and conference proceedings. His research interests are in the area of mobile computing, routing protocols, fault-tolerant computing, and interconnection networks. He served as a program vice chair for 2000 International Conference on Parallel Processing (ICPP) and a program vice chair for 2001 IEEE International Conference on Distributed Computing Systems (ICDCS). He is a program co-chair for the IEEE 1st International Conference on Mobile Ad-hoc and Sensor Systems (MASSÕ04). He was a co-guest-editor of a special issue in IEEE Computer on Ad Hoc Networks. He also edited several special issues in Journal Parallel and Distributing Computing (JPDC) and IEEE Transactions on Parallel and Distributed Systems (TPDS). He is the author of the text Distributed System Design published by the CRC Press. Currently, he serves as an Associated Editor in IEEE Transactions on Parallel and Distributed Systems and three other international journals. He is a recipient of the 1996 1997 and 2001 2002 Researcher of the Year Award at Florida Atlantic University. He served as an IEEE Computer Society Distinguished Visitor. He is a Member of ACM and a Senior Member of IEEE.